Beruflich Dokumente
Kultur Dokumente
5, SEPTEMBER 2011
599
I. INTRODUCTION
OBUST and reliable identification of individuals is clearly
a most important tool in, for example, antiterrorism scenarios, yet many traditional approaches to identity determination
and people monitoring have manifest flaws and weaknesses.
Biometrics-based approaches provide a potentially very powerful option, providing both the possibility of confident identity prediction in many cases, while also allowing a lower
level monitoring function based on probabilistic assessments
or confidence-based predictions in other circumstances.
Although formal biometric identification systems are not
guaranteed in all circumstances to be effective in protecting
against terrorist threat, and although establishing identity unequivocally is always challenging, a principal benefit of biometric technologies is their ability to bind activity to an individual, thereby increasing the challenges associated with assuming
different identities, or masking identity across different activities, across physical boundaries, and even across international
borders [1].
Manuscript received December 28, 2009; revised April 12, 2010; accepted
June 26, 2010. Date of publication August 3, 2010; date of current version
August 19, 2011. The work of M. C. C. Abreu was supported by the Coordenaca o
de Aperfeicoamento de Pessoal de Nvel Superior (CAPES) (Brazilian Funding
Agency) under Grant BEX4903-06-4. This paper was recommended by Associate Editor J. Tang.
The authors are with Department of Electronics, University of Kent,
Canterbury, Kent CT2 7NT, U.K. (e-mail: mcda2@kent.ac.uk; m.c.fairhurst@
kent.ac.uk).
Digital Object Identifier 10.1109/TSMCC.2010.2056920
However, there is an argument to be made that it is necessary to be more imaginative about how to deploy biometric
identification processing if the full potential of these technologies is to be most effectively harnessed in this type of context.
For example, it is important to acknowledge that antiterrorist
applications are by their very nature rarely well defined, and
identifying, tracking, and predicting the behavior of individuals intent on avoiding detection is a major challenge; however,
issues of identity are approached. One way in which creatively
to enhance the value of the processing methodologies, which
underpin biometric identification systems is to recognise that in
many practical situations identity evidence in relation to any individual is likely to be fragmented and distributed, and thus, no
single-simple recognition algorithm is likely to be completely
effective. It may, therefore, be advantageous to shift the focus
from identification in an absolute sense to a process, which
evaluates multiple sources of identification evidence in order to
predict, with a greater degree of confidence than would otherwise be possible, the identity of an individual, whose activities
are being monitored [2].
Biometrics is now a very well established and intensively researched subject area, yet still may be considered to represent a
group of technologies, which are at a pivotal stage of development [3]. Despite the fact that the biometrics field has reached a
level of maturity, which can now support reliable practical applications (for example, in relation to travel documents, managing
worker movements in the construction industry, etc.), there are
still some important unresolved issues, which need to be addressed and, indeed, the search for improved performance is a
continuing issue [4].
In most applications, however, increasing the accuracy with
which a biometric solution can identify an individual or verify claimed identity, is a very important issue, especially since
the search for improved accuracy also raises related questions
about the possibility of tradeoffs between different types of
error, how to configure the underpinning processing systems,
and the implications across a range of performance factors of
structuring an identification system in a particular way. Approaches to improving accuracy have focused principally on
optimizing processing within chosen individual modalities (increasing the reliability of minutiae extraction from fingerprint
images [5], for example, or developing alternative face matching algorithms [6]), extending existing modalities (for example,
moving to 3-D facial imaging [7]), or seeking greater optimization of the generic classification engine deployed [8]). Beyond
this, there has been an increasing interest in acquiring better and
600
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 41, NO. 5, SEPTEMBER 2011
Fig. 1.
601
TABLE I
CONFIDENCES OF THE SAMPLE FOR IDENTITY AND GENDER PREDICTION
3) When the identity prediction classifiers disagree, the identity output of the system is the user identity, which has
received most votes from the soft-biometric prediction
classifiers and the identity prediction classifiers.
4) When there is a tie, the identity prediction classifier with
the greatest confidence provides the output of the system.
As an example to illustrate the operation of this method,
a hypothetical three-user (A, B, and C) identification task is
considered, the inputs of which contain three features (fea1,
fea2, and fea3). The proposed fusion system is composed of
two identity prediction classifiers (Cl-I-1 and Cl-I-2) and two
gender prediction classifiers (Cl-G-1 and Cl-G-2). The gender
prediction classifiers can generate either male or female as
an output. After the training process for these classifiers, the
following hypothetical test pattern is presented: fea1: 0.7; fea2:
0.4; and fea3: 0.22. User A is male, and users B and C are
female. The classifier outputs for this imagined situation can
be seen in Table I, where the winner class for each classifier is
italicized.
In this case, the following output will be generated by the
system.
1) User A: one vote (Cl-I-1).
2) User B: three votes (Cl-I-2, Cl-G-1, and Cl-G-2).
3) User C: no vote.
The predicted identity of the system would be user B, because
it has the majority of votes.
B. Sum-Based Fusion Method
Sum-based fusion [14] is a linear fusion-based method that
takes into account the confidence degree for each class of each
classifier. In this sense, when an input pattern is presented to the
base classifiers, the degrees of confidence for each class output
are added to the other related outputs giving a score to that class.
The winner class, and hence, the identity label of the system is
the class with the highest score.
The same general system structure as for the majority voting
approach is used, as shown in Fig. 1. We also need to adapt this
method to our current purpose as follows.
1) All the classifiers generate their output for the given input
sample.
2) The corresponding class confidence is added in a class
score.
3) In the event of a tie, the identity prediction classifier with
the greatest confidence provides the output of the system.
Fig. 2.
602
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 41, NO. 5, SEPTEMBER 2011
Puni =
(1)
where
1) Di is the difference between the current i feature of the
test pattern and its training mean for the identity classifier;
2) Dj,i is the difference between the current i feature for the
current winner class j of the test pattern and its training
mean for the soft-biometric classifier;
3) Si is the sensitivity of the classifier to the corresponding i
feature of the chosen class for the identity classifier;
TABLE II
SENSITIVITY ANALYSIS OF THE PREDICTION CLASSIFIERS
TABLE III
TRAINING MEAN OF ALL FEATURES
TABLE IV
ABSOLUTE DISTANCE OF THE TEST PATTERN AND ITS TRAINING MEAN
603
604
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 41, NO. 5, SEPTEMBER 2011
TABLE V
ERROR MEAN AND STANDARD DEVIATION OF ALL INDIVIDUAL PREDICTION
CLASSIFIERS USED WITH THE FINGERPRINT AND FACE DATABASES
subsets. Nine out of ten of the training subsets are used to train
the classifier and the tenth subset is used as the test set producing
an error measurement each time. The procedure is repeated ten
times, with a different subset being used as the test set in each
case and the average of these ten errors represents the error rate
for that classifier.
The comparison of two classification methods is accomplished by analyzing the statistical significance of the difference
between the error mean of the classification rate on independent
test sets from the methods evaluated. In order to evaluate this
the p-value provided by the t-test [32] measures the degree of
confidence in the result. In our case, we use a confidence level
of 95%, where one sample is deemed to be statistically different
from another only when the p-value is lower than 0.05.
V. RESULTS AND DISCUSSION
In order to interpret and evaluate the results of the fusion process, it is important first to analyze the results of the individual
classifiers that will be used as the base elements in the fusionbased structure. Table V shows the results obtained using all
seven classifiers for each prediction task.
It can be seen in Table V that the identity prediction classifiers
generate a similar error mean to that returned with the softbiometric prediction process. This is not expected because in
the identification task each user identity is considered to be
a class, when in, for example, the age prediction process the
classifier only needs to give a prediction based on a choice
from just three classes, which makes the classifier task much
less challenging. Moreover, this result also demonstrates that
soft-biometric prediction can be effectively realized.
Because our aim in this paper is to predict identity, it is
important to know if there is any statistical difference between
the identity prediction classifiers. According to the interpretation
of the t-test and comparing the best classifier (SVM) with the
others, it can be asserted that the SVM classifier is statistically
better than all other classifiers apart from the FMLP for the two
databases with 0.36 and 0.39 p-values for face and fingerprint,
respectively.
605
TABLE VI
ERROR MEAN AND STANDARD DEVIATION OF ALL FUSION AND THE
NEGOTIATION METHOD USED WITH THE FACE AND FINGERPRINT DATABASES
The soft-biometric prediction classifiers presented, in general, a good level of accuracy. In the fingerprint database, the
age prediction classifiers generally return the best accuracies despite having to choose among three classes. On the other hand,
the gender prediction classifiers perform better than the age prediction classifiers in the face database. This suggests that the
age effects in fingerprint are greater than the difference between
the genders and the difference between the genders has a greater
impact for the face database than when age is considered.
Overall, the results returned by the individual classifiers are
rather typical of others reported in the literature, but the fusion
methods proposed are specifically aimed at trying to combine the
advantages of different classifiers and improve overall accuracy.
In our experimental work, we used three classifiers in each
prediction task considered. The particular structures considered
are as follows:
1) Identity (MLP/Jrip/SVM) + Age (FMLP/DT/KNN) +
Gender (RBF/Jrip/SVM);
2) Identity (FMLP/DT/KNN) + Age (RBF/KNN/SVM) +
Gender (FMLP/SVM/Jrip);
3) Identity (RBF/KNN/SVM) + Age (FMLP/SVM/Jrip) +
Gender (MLP/Jrip/SVM);
4) Identity (FMLP/SVM/Jrip) + Age (MLP/Jrip/SVM) +
Gender (RBF/KNN/SVM).
Table VI shows the accuracy achieved by the fusion (voting and sum) and the negotiation (sensitivity) methods. We
used a value of ten for both the constants (Ca and Cm b ) in
the negotiation method, chosen to provide a strong punishment
level, which this constant controls. The fusion methods show
an improvement of 50% on average when compared with the
individual classifiers, which is clearly a significant gain in performance. However, the sensitivity-based method achieved on
average an 80% improvement when compared with the individual base classifiers, showing a substantial further performance
enhancement.
According to the outcome of the t-test, all these combination
methods are statistically better than all the individual classifiers.
This fact confirms that there are advantages in combining different methods to improve accuracy. Also, it is very interesting to
see that the sensitivity-based method is always statistically better
than the two centralized fusion-based fusion methods. Again,
Fig. 3.
606
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 41, NO. 5, SEPTEMBER 2011
607
Michael Fairhurst received the B.Sc. and Ph.D. degrees in 1969 and 1972, respectively.
He is currently in the Department of Electronics,
University of Kent, Canterbury, U.K. He has authored
or coauthored more than 350 papers in the scientific literature. His research interests include computational architectures and algorithms for image analysis and classification, and applications, including
handwritten text reading and document processing,
medical image analysis and, especially, security and
biometrics.
Prof. Fairhurst has been a member of numerous conference, workshop, and
government-sponsored committees, and sits on the Editorial Boards of several
international journals. He is an Elected Fellow of the International Association
for Pattern Recognition in recognition of his contributions to the field.